How to Prompt an AI Agent - The Seven-Line Brief

Prompting a chatbot is asking a question. Prompting an agent is handing over a job. The difference is the difference between asking a colleague what time the post goes and asking them to sort the post for the week.

An agent can open files, search the web, click through sites, write code and send things, usually in a long run of steps you do not watch. That makes the prompt less like a question and more like a Monday-morning brief to a new starter you cannot interrupt. Here is how to write one that does not end in tears.

What is an AI agent prompt?

An AI agent prompt is a set of instructions that tells an AI system what outcome to achieve, which tools and information it may use, what it must not do, and when to stop and check with you. It differs from a chatbot prompt because the AI will take many actions on its own before it answers.

If the difference between agents and chatbots is still fuzzy, read AI agents vs chatbots first, then come back.

How is an agent prompt different from a chatbot prompt?

What should an agent prompt include?

Seven lines. Skip one and the agent fills the gap with a guess, and agents guess confidently.

1. GoalThe outcome in one sentence. Not the method.
2. Done whenHow the agent knows it has finished. A number, a file, a list of ten.
3. Tools and accessWhat it may use, and what it may read or change.
4. BoundariesWhat it must never do. Name the expensive mistakes.
5. Check-insWhere it must stop and ask before carrying on.
6. OutputWhat you want back, in what shape, saved where.
7. If stuckWhat to do when it cannot continue. Usually: stop and tell you.

GOAL: Find ten UK accountancy firms with 5 to 20 staff within 30 miles of Oxford that have posted about AI in the last 90 days.

DONE WHEN: I have a spreadsheet of exactly ten firms, each with a source link and a one-line reason it qualifies.

TOOLS: Web search and the browser. Read-only. Do not sign up to anything or fill in any forms.

BOUNDARIES: Do not contact anyone. Do not include firms you cannot verify from a public page. Do not guess.

CHECK-INS: After the first three firms, show me your method and wait for my OK.

OUTPUT: A table with firm, town, staff estimate, source link, reason. British English.

IF STUCK: Stop and tell me what is blocking you. Do not work around it.

How much freedom should I give an agent?

Think of a leash with four lengths. Start short. Lengthen it only when the agent has earned it on that specific task.

What do real agent prompts look like?

A Monday-morning competitor watch (level 1)

Every Monday, check the websites of these five competitors [list]. Look only for new pages, price changes and new case studies since last Monday. Report what changed in a table of competitor, what changed, link, why it might matter. If nothing changed, say so in one line. Do not visit any other sites.

A tidy-up job for a shared folder (level 2)

Look through the folder 'Client Documents 2026'. Propose a new folder structure and a renaming scheme, but do not move or rename anything. Save your proposal as proposal.md in the same folder. List any files you are unsure how to classify.

A first project with a coding agent (level 3)

Coding agents such as Claude Code are the most capable agents most people will meet. They also show why the brief matters. The more precise the goal and the stricter the boundaries, the better the result.

Build a single-page website for a dog groomer in Witney with a price list, opening hours and a contact form that emails me. Plain HTML and CSS only. Before writing any code, show me your plan and wait for approval. Do not install anything or change any files outside this folder. When finished, tell me what to do to publish it.

If that last example made you curious, our Introduction to Claude Code and Building Agents in Claude Code courses are written for people who are not developers.

How do I check what an agent did?

Reading the final answer is not enough. An agent can reach the right answer by a dubious route or report success on something it did not finish. Ask for a debrief and read the evidence.

Give me a debrief on what you just did. List each step you took, each source you used, anything you could not verify, and anything you assumed. Then tell me what you would check next if this were your job on the line.

How do AI agents go wrong?

• They solve the wrong problem confidently. Fix with a clear done-when and an early check-in.
• They carry on after something breaks. Fix with an if-stuck line that says stop.
• They take the shortcut you did not rule out. Fix by naming the expensive mistakes in the boundaries.
• They treat web pages as instructions. A hostile page can contain text aimed at the agent. Keep agents away from untrusted content when they hold real access.

How do I start with agents?

1. Pick a boring taskSomething repetitive, low stakes and easy to check.
2. Run it at level 1Let the agent look and report. Compare with what you would have done.
3. Tighten the briefEvery surprise in the result is a missing line in the prompt. Add it.
4. Lengthen the leashMove to drafts, then to approvals, once it behaves three times running.
5. Write it downSave the final brief. The next agent job starts from it.

Learn to brief an agent before one briefs itself

Our agent courses take you from your first safe, read-only agent to building your own in Claude Code, with plenty of time for the questions people are usually too polite to ask.

Building Agents in Claude CodeTalk to Harry

• Introduction to Claude Code Your first coding agent.
• Advanced Claude Code Skills, context and agent workflows.
• AI Policy Workshop Rules for what your agents may touch.
• Building Agents in Claude Code Design, test and supervise your first agents.

Harry Lang
Harry Lang founded The Oxford AI School in Witney, Oxfordshire. He trains UK businesses and teams to use AI well, writes the Marketing Unfiltered newsletter, and is the author of Learn To Fly With AI.